Nearly a Year After Bubble Duck Shut Down, I Dug Up My Old Usage Records

I compared Bubble Duck with the general conversational models I usually use, actually ran it, over more than a month. The time was before it shut down, last autumn. By the time I wrote this, it was no longer downloadable, reportedly ceased operations, so this is only a retrospective, not a recommendation.

Bubble Duck was a C-end product by StepFun, officially called "AI Open World," where a large model drives a bunch of characters, you chat with them, advance the plot, make choices, and can also create your own characters and plots. Here, a character is a preset persona with personality and background; every line you say is a dialogue with that persona. Open world means many storylines are laid out for you to choose from, not like a map in a game.

First glance inside: the interface shows a row of character cards, click one and start chatting. Its interaction design is quite convenient—plot advancement is just clicking continue, or picking one of several options, no need to know how to write prompts.

Some reviews say that even after many rounds of interaction and several chapters, the plot logic remains complete, with almost no contradictions or forgetting.

I verified this. I chatted with one character up to chapter ten-something, came back two days later, and asked about a setting I had defined in chapter one—it picked it up. This was its strongest point at the time: super-long context memory plus real-time web search, long conversations don't collapse. I tried doing the same with a general model, and later in the chat it started forgetting, persona drifting, and I had to repeatedly paste the settings back.

Now the surprises are done, onto the sticking points.

First, the character creation form is more tiring to fill out than expected. You need to write persona, background, speaking style, opening line; without a template it's easy to write a mess. I created three, only one could run smoothly; the other two after a few rounds reverted to a general assistant with an AI tone. Maybe my prompts were bad, but the barrier is indeed there.

Second, web search in the plot is a bit immersion-breaking. It can pull in real-world information, but you want immersion; once real-world info comes in, the immersion breaks. Later I basically treated this capability as decoration.

Third, and what I as an investor care about most, it couldn't get users to keep paying. The model capability was solid, but why users stayed was due to characters and content, and neither became platform assets. Characters users created themselves stick to the users, not the platform.

According to Qimai data, Bubble Duck's downloads across all platforms exceeded 6 million. In the same period, Xiaoice's XEVA downloads exceeded 3 million and also shut down around the same time.

6 million downloads isn't small, but downloads are customer acquisition cost, not a moat. Where are the people truly willing to pay for companionship? Look at the neighboring category for clarity.

"Love and Deepspace" had a single-week revenue of nearly 89 million RMB in week 21 of 2025.

It's a game, not an AI product, but it understands companionship better than this batch of AI companions. Characters have art, voice acting, and a plot team continuously supplying content; what users pay for is the certainty of that content. On Bubble Duck's side, the model is general capability, content is user-generated, and the platform didn't hold the supply side.

Then came the path: in June 2025, large-scale investment stopped, the team merged into StepFun AI, a shutdown notice was issued in September, and later it officially ceased operations. When a model company does C-end applications, the model is strong, but content operations and IP aren't things model capability can casually solve.

My judgment is it depends, and it depends for specific people.

If you're doing AI companionship or character products, I suggest dissecting Bubble Duck as a sample, focusing on what it got right with long memory and why it still couldn't retain users. Long memory is an entry ticket, not a barrier—any replicable technology isn't a barrier.

If you're looking to invest in this direction, don't value it by model capability; look at three things: is there character IP, is there continuous content supply, and are paying users paying for characters or for features. If any one of the three doesn't hold, the valuation logic collapses.

Ordinary users have nothing to try now; it's already shut down. Similar products are still running; when you pick one, you can first ask: the characters inside, are they raised by the platform itself, or by the users themselves?